Of the 71 data systems considered, 53 data systems were outside the study scope based on exhibiting one or more of the following characteristics: mechanism-specific (24 data systems); violence-only (10); not routinely released, have been discontinued or are one-time surveys (8); fatal-only (8); no injury-related variables (5); occupational-specific (3); body region-specific (2); not publicly available (2), and/or no injury mechanism variables (1). Ultimately, 18 data systems total were included (online supplemental appendix 1):

When selecting a data system to use for analyses, it is key to know data system characteristics in detail. A few specific examples of data system characteristics and changes over time are highlighted further; however, this list is not exhaustive. Data system documentation should be consulted prior to performing analyses.

A data system’s sample and injury case definition both impact the types of analyses that may be appropriate for a given data system and may make comparisons across different data systems challenging. Each data system collects data through a variety of methods (eg, probability sample, convenience sample and census) and are abstracted from diverse sources (eg, ED visits, hospitalisations, trauma admissions, hospital readmissions, office-based physician visits, household interviews, EMS activations, ambulatory surgery and medical claims). Similarly, since injury case definitions rely on the coding system from which they are derived, injury case definitions may also vary across data systems, often causing incomparability between injury estimates across data systems.

Additionally, injury coding can change over time both between and within data systems. For example, data systems that rely on the ICD-9-CM/ICD-10-CM coding systems for injury coding have been affected by the transition from the 9th revision to the 10th revision on 1 October 2015.^{27 28} There are also slight coding changes within these coding systems each October when codes may be added, modified or deleted.²³ The deletion or modification of codes may render trend analyses challenging over certain time frames for specific codes within a data system. Analysing trends over the 2015 change for data systems that use ICD-9-CM/ICD-10-CM coding should be done with caution.

Some data systems do not rely on the ICD coding system to define injury mechanism, and thus injury mechanism in those systems may contribute different or additional information. For example, NHIS contains injury risk factor variables, such as a variable indicating whether or not an individual was wearing a helmet or seatbelt at the time of a crash,²⁹ and this information does not have a comparable ICD-9-CM/ICD-10-CM code. NEISS and NEISS-AIP include narratives describing the context of the injury. The context of an injury, whether assessing fatal or non-fatal injury, is often lacking in data systems, especially claims data, but is nonetheless a deeply important aspect of understanding how to prevent injuries.³⁰

Sampling framework changes and redesigns to data elements over time may also make longitudinal analyses over certain periods complicated. For example, prior to 2012, NIS captured all discharges from a sample of hospitals, used the American Hospital Association hospital and discharge definition and included data elements that were not necessarily supplied by all states. In 2012, NIS’s sampling framework was redesigned to capture a sample of discharges, to use the state organisation’s hospital and discharge definition, and to use a standard set of data elements. This redesign was carried out to minimise sampling error, improve confidentiality and produce more precise national estimates. The data user should be cautious about performing NIS trend analyses with data before and after 2012 and be certain to use the proper trend weights as described in the NIS documentation.³¹ Other data systems have also experienced redesigns or have undergone revisions to data elements. NHIS redesigns its sampling process following each decennial census.²⁵ OASIS redesigned their data most recently in 2019.¹² Modifications to data elements and redesigns of sampling frameworks should be identified prior to initiating any longitudinal analyses.

Non-fatal injury research questions, study plans and analyses can vary widely; goals for injury surveillance can also vary. As mentioned previously, NEISS and NEISS-AIP provide important details about the context in which the injury occurred, which can be helpful for surveillance and research activities. Narratives in NEISS-AIP were used in a surveillance study to categorise traumatic brain injuries (TBIs) based on the sports-related consumer product involved and to understand the context surrounding the TBI occurrence.³² Haarbauer-Krupa et al³³ used NEISS and NEISS-AIP to categorise unintentional TBIs among children, indicating any product and the activity associated with the fall. Using the narratives, certain action words were selected (eg, ‘jump’ and ‘climb’) that showed child-initiated action that precipitated a fall and subsequent brain injury, and these cases were compared with cases where a caregiver caused the precipitation of the event or series of events leading to the TBI injury. NEISS-AIP and NEISS are appropriate systems for this research, considering the amount of detail needed from narratives to categorise the injuries, which other data systems, such as the ICD-9-CM/ICD-10-CM-based data systems, do not supply.

Using a combination of data systems with different samples can help provide a more comprehensive picture of injury burden. Peterson et al³⁴ investigated the average medical cost of fatal and non-fatal injuries by type using three data systems that used data from three different sources (medical claims (MarketScan), ED visits (NEDS) and hospitalisations (NIS)). This resulted in capturing a larger range of medical costs due to injury. The study period avoided the ICD-9-CM/ICD-10-CM transition date (1 October 2015). This avoided potential inconsistency in injury definition/coding that could arise by analysing a study population containing cases coded using both the ICD-9-CM and ICD-10-CM coding systems.

This study provides a high-level overview of data systems commonly used for non-fatal injury analysis and can be a resource when seeking to address surveillance or research questions. However, it represents a starting point and does not include all details necessary to determine the most appropriate data system to use. Prior to any analyses, documentation for each data system should be carefully investigated in order to understand other possible limitations, any changes which may have occurred within the time frame of interest, and the data system’s suitability for the question being asked. Data quality, not directly addressed in this study, is also important to account for; the reliability of results from a well-planned analysis relies heavily on a good quality data source.

Additionally, many of the data system characteristics were captured for a single data year, and these characteristics could be different for other data years. Further, the data systems included in this paper are not an exhaustive list of commonly used data systems for injury analysis. Fatal-only data systems; violence-specific, topic-specific, risk factor-specific or occupational-specific data systems; and one-time surveys were not included (online supplemental appendix 2). Data systems in online supplemental appendix 2 may be a better source of data for specific injury mechanism surveillance and research questions and can provide additional context to help inform decision-making. For example, the Fatality Analysis Reporting System covers only motor vehicle traffic crash fatalities but does so with a wider range of variables than a claims-based data system.